diff --git a/.DS_Store b/.DS_Store index b18e10abfd6fc0ca82bede3bb58e11fccb245296..7015d2de83655bf366d5115e417de29fbca8abbd 100644 Binary files a/.DS_Store and b/.DS_Store differ diff --git a/manuscript/RSF.pdf b/manuscript/RSF.pdf index cc8d442b2758e206e42154412434900d25379acd..8e5468435c78419fea3b0fe7d1d04e81bf9a7895 100644 Binary files a/manuscript/RSF.pdf and b/manuscript/RSF.pdf differ diff --git a/manuscript/RSF.tex b/manuscript/RSF.tex index 4ad2ad792572fc941c6f6461e9b1cef6cfa96b99..5f0a7ef1f5894610c9c636cd449ca99980359de9 100644 --- a/manuscript/RSF.tex +++ b/manuscript/RSF.tex @@ -96,12 +96,11 @@ Key recent deep learning papers [Schuman et al.]. Other learning approaches [NNM ------- The past few decades have ushered in revolutionary advancements in medical imaging which have helped elucidate relationships between biological structures and physiological function in health and disease. Combining computer-aided diagnosis with imaging data further augments physicians’ ability to make higher precision clinical decisions. Perhaps no better example of this could be made than with the impact that optical coherence tomography (OCT) has had on eye care. OCT is a non-invasive diagnostic imaging tool which employs principles of optical interferometry using a low-coherence light source to obtain cross-sectional digitally reconstructed images of biological tissue [1]. Often described as an optical analog to ultrasound, OCT images are generated from detection of electrical field produced from the echoed tissue and reference signals [Fujimoto 2016]. OCT datasets are comprised of multiple A-scans acquired in rapid succession to form cross-sectional images, or B-Scans, of the subject tissue, which can also be acquired in rapid succession to obtain a cube image. Current generations of spectral domain (SD) OCT are capable of acquiring an axial resolution of 5-8 microns with a lateral resolution of 6-20 microns and a scanning rate of 100,000 A-scans per second, making them ideal for Capture of 3D volumetric data in vivo. [Alexopoulos, et al 2022]. - Although predominantly used in clinical eye practice, OCT use spans beyond ophthalmology. OCT has clinical applications in cardiology, otology, dermatology, and dentistry [Ali 2022]. OCT’s ability to penetrate up to 3mm of tissue and produce high resolution images have provided use cases in diagnosing myocardial infarction with non-obtrusive coronary arteries [Reynolds, 2021], facilitating cochlear implants surgery [Starovoyt, 2019], diagnosing basal cell carcinoma [Chen, 2021], and even dental cavity detection [Hariri, 2013]. More recently, the sub-field of ophthalmic imaging known as Oculomics has shown the capability of using OCT to make prognostications of non-ocular systemic disorders like cardiovascular [Chan 2023] and neurological disease [Suh 2023, Lin 2024] through ocular imaging, and even improvements on estimating phenotypic age for predicting mortality [Nusinovici 2022]. - This ability to safely and quickly acquire ocular images in a longitudinal and reproducible manner, has revolutionized the way clinicians diagnose and manage blindness causing diseases such as glaucoma. Glaucoma is the global leader of irreversible blindness with a prevalence 3-5% among individuals aged 40 and older, with projections indicating that the number of affected individuals could rise to 112 million by 2040 due to population aging [Tham 2014]. Glaucoma is often asymptomatic until moderate to severe stages of disease, which is why OCT play’s a critical role in mitigating disease through early detection of retina structural changes. However, OCT alone is insufficient to determine disease severity. Glaucoma standard of care involves ophthalmic imaging of specific retinal features in conjunction with functional assessment of the patient’s visual field (VF) through standardized automated perimetry [Zhiqi 2024]. One of the prominent ways to categorize disease severity and visual field defect is through visual field mean deviation (VFMD), which is a numerical estimation of light in decibels (dB) that an individual eye can perceive when compared to an age-matched normative database [Thirunavukarasu, 2024]. Structural changes in the retina are not linearly correlated with VFMD. Broken stick statistical models show little to no correlation between mean RNFL threshold for VFMD until a tipping point is reached and statistically significant associations are observed [Wollstein, 2012]. +Artificial intelligence (AI) applications using deep learners (DL) to detect, diagnose and predict disease progression have a potentially enormous impact on public health [Schuman 2022]. The most common approaches look to classify glaucoma based on clinically accepted structural features such as the retinal nerve fiber layer (RNFL), ganglion cell inner plexiform layer (GCIPL) and optic nerve head (ONH) [Prahs 2018, Grewal 2008, Christopher 2018, Shi 2024, Berenguer 2021]. Furthermore, incorporating features such as Ganglion cell complex (GCC) thickness, ONH macrostructure, and RNFL reflectance maps into DL model training significantly improves diagnostic accuracy [Tan 2024]. Although there is no consensus on the exact clinical classification of glaucoma, the structure function relationship between OCT and VF are critical in understanding disease severity and progression. However, VF testing can be challenging for patients to perform accurately. The test which can take anywhere between 4-8 minutes per eye, requires users to sit still and fixate on singular point in a testing bowl and click a button anytime a point of light is perceived in their field of vision. Test difficulty, combined with an older patient demographic makes the test more variable than OCT. Therefore, considerable efforts have been made towards predicting VF outcomes based on OCT. Previous studies have shown the ability to derive Spatial relationship between structure and function using OCT and VF point-by-point estimation [Zhiqi 2024] based on stochastic optimization DL models [Kingma 2014]. +Unsupervised feature agnostic approaches using OCT have shown promise in predicting disease progression measured by VF. Chen et al. [Zhiqi 2023], describe a 3D based ResNet18 Convolutional neural network (CNN) capable of inferring pointwise VF sensitivities directly from segmentation-free OCT 3D volumes. Similarly, attention-guided network approaches have demonstrated improved glaucoma detection by taking OCT volumes and computing dual 3D gradient class activation heatmaps to predict the Visual Field Index (VFI) estimation [George 2020]. DL approaches using OCT volumes from glaucoma subjects have identified 14 non-clinically defined surface shape patterns near the ONH which capable of predicting specific VFMD loss rate with a best coefficient of determination of r2=0.37 [Saini 2022]. These clinically agnostic DL approaches using OCT volumes have demonstrated the ability to identify previously undiscovered biomarkers for improving the prediction of glaucomatous functional defects, but, perhaps due to the insufficient confidence indices or inability to explicitly and metrically define the relationships, remain limited to research applications. -Artificial intelligence (AI) applications for detecting and diagnosing glaucoma and predicting disease progression have enormous potential impact on public health [Schuman 2022]. The most common approaches look to classify glaucoma based on widely clinically accepted structural features such as the retinal nerve fiber layer (RNFL), ganglion cell inner plexiform layer (GCIPL) and optic nerve head (ONH) [Prahs 2018, Grewal 2008, Christopher 2018, Shi 2024, Berenguer 2021]. Combining additional features such as GCIPL thickness, and ONH macrostructure, and RNFL reflectance maps combined with DL models significantly improve diagnostic accuracy [Tan 2024]. Although there is no consensus on the exact clinical classification of glaucoma, the structure function relationship between OCT and VF are considered the gold standard. However, VF testing can be challenging for patients to perform accurately making the test results more reliable. Therefore, many efforts have been made towards predicting VF outcomes from OCT. Spatial relationship between structure and function were derived using OCT and VF point-by-point estimation [Zhiqi 2024] based on stochastic optimization models [Kingma 2014]. Feature free approaches are another approach which have shown promise in predicting disease progression. 3D based ResNet18 Convolutional neural network (CNN) was capable of inferring pointwise VF sensitivities directly from segmentation-free OCT 3D volumes [Zhiqi 2023]. Clinical feature agnostic approaches using OCT volumes are ideal as they may elucidate undiscovered biomarkers for predicting glaucoma. Attention-guided network approaches have shown the ability to improve glaucoma detection [George 2020]. Recent studies using OCT volumes from glaucoma subjects centered on the ONH have identified 14 surface shape patterns capable of improving prediction of VFMD [Saini 2024]. -------